Fraud anomaly detection on transaction streams

Project Description

We review flagged transactions manually against rules written two years ago. Our false positive rate is high enough that the team has started rubber-stamping. We want anomaly detection that reduces the review queue without missing the cases that matter. Precision and recall trade off here and we want to make that trade deliberately, with numbers. We can share a labelled set of confirmed fraud, though it is small — how you handle the class imbalance is a big part of what we are evaluating.

Skills Required